Computerized Medical Imaging and Graphics最新文献

{"title":"Multimodal data fusion with irregular PSA kinetics for automated prostate cancer grading","authors":"Oleksii Bashkanov ,&nbsp;Lucas Engelage ,&nbsp;Niklas Behnel ,&nbsp;Paul Ehrlich ,&nbsp;Christian Hansen ,&nbsp;Marko Rak","doi":"10.1016/j.compmedimag.2025.102625","DOIUrl":"10.1016/j.compmedimag.2025.102625","url":null,"abstract":"<div><div>Prostate cancer (PCa) detection and accurate grading remain critical challenges in medical diagnostics. While deep learning has shown promise in medical image analysis, existing computer-aided diagnosis approaches primarily focus on image recognition, overlooking patient-relevant information. Additionally, current multimodal fusion approaches face a significant limitation in their inability to effectively integrate and analyze irregular time-series data, such as prostate-specific antigen (PSA) measurements alongside imaging data, particularly in cases where measurements are taken at inconsistent intervals. Here, we present a novel multimodal fusion framework that effectively combines imaging data with longitudinal patient information, including irregular PSA measurements, demographic data, and laboratory results. Our architecture employs a custom embedding technique to handle temporal sequences without requiring complex preprocessing or imputation steps. We evaluated our framework on a comprehensive dataset of prostate cancer patients from multiple clinical centers, encompassing both internal and external validation cohorts. The integration of temporal PSA information with imaging embeddings resulted in superior performance compared to traditional image-only approaches, demonstrating an improved area under the receiver operating characteristic curve (AUC) (0.843 vs. 0.808) for detecting clinically significant prostate cancer (csPCa). Our approach also achieved substantially more accurate prostate disease grading with a quadratic weighted kappa (0.645 vs. 0.557), validated on 630 cases from the same institution. The model demonstrated robust performance (AUC of 0.765) when evaluated on an external dataset comprising 419 cases from multiple European centers, utilizing 160 different MRI devices. When compared to experienced radiologists using PI-RADS scoring, our model showed higher sensitivity (74.5% vs.<!--> <!-->62.2%) at matched specificity (76.5%) while maintaining comparable performance (98.3% vs.<!--> <!-->98.1%) at high-sensitivity operating point. Our approach shows particular promise in reducing unnecessary biopsies while maintaining high detection sensitivity, suggesting significant potential as a clinical decision support tool.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"124 ","pages":"Article 102625"},"PeriodicalIF":4.9,"publicationDate":"2025-08-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144852134","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"LR-COBRAS: A logic reasoning-driven interactive medical image data annotation algorithm","authors":"Ning Zhou,&nbsp;Jiawei Cao","doi":"10.1016/j.compmedimag.2025.102623","DOIUrl":"10.1016/j.compmedimag.2025.102623","url":null,"abstract":"<div><div>The volume of image data generated in the medical field is continuously increasing. Manual annotation is both costly and prone to human error. Additionally, deep learning-based medical image algorithms rely on large, accurately annotated training datasets, which are expensive to produce and often result in instability. This study introduces LR-COBRAS, an interactive computer-aided data annotation algorithm designed for medical experts. LR-COBRAS aims to assist healthcare professionals in achieving more precise annotation outcomes through interactive processes, thereby optimizing medical image annotation tasks. The algorithm enhances must-link and cannot-link constraints during interactions through a logic reasoning module. It automatically generates potential constraint relationships, reducing the frequency of user interactions and improving clustering accuracy. By utilizing rules such as symmetry, transitivity, and consistency, LR-COBRAS effectively balances automation with clinical relevance. Experimental results based on the MedMNIST＋ dataset and ChestX-ray8 dataset demonstrate that LR-COBRAS significantly outperforms existing methods in clustering accuracy, efficiency, and interactive burden, showcasing superior robustness and applicability. This algorithm provides a novel solution for intelligent medical image analysis. The source code for our implementation is available on <span><span>https://github.com/cjw-bbxc/MILR-COBRAS</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"124 ","pages":"Article 102623"},"PeriodicalIF":4.9,"publicationDate":"2025-08-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144831353","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"TG-Mamba: Leveraging text guidance for predicting tumor mutation burden in lung cancer","authors":"Chunlin Yu,&nbsp;Xiangfu Meng,&nbsp;Yinhao Li,&nbsp;Zheng Zhao,&nbsp;Yongqin Zhang","doi":"10.1016/j.compmedimag.2025.102626","DOIUrl":"10.1016/j.compmedimag.2025.102626","url":null,"abstract":"<div><div>Tumor mutation burden (TMB) is a crucial biomarker for predicting the response of lung cancer patients to immunotherapy. Traditionally, TMB is quantified through whole-exome sequencing (WES), but the high costs and time requirements of WES limit its widespread clinical use. To address this, we propose a deep learning model named TG-Mamba, capable of rapidly predicting TMB levels based on patients’ histopathological images and clinical information, and further estimating specific TMB values. Specifically, we employ a parallel feature extraction strategy. The upper layer consists of a series of text-guided attention modules designed to extract diagnostic textual features. Meanwhile, the lower layer leverages the VMamba backbone network for image feature extraction. To enhance performance, we design a novel hybrid module, Conv-SSM, which combines convolutional layers for local feature extraction with a state-space model (SSM) to capture global dependencies. During the feature extraction process, textual features progressively guide the extraction of image features, ensuring their effective integration. In a cohort of non-training lung cancer patients, TG-Mamba achieved an area under the receiver operating characteristic curve (AUC) of 0.994 in classification tasks and a mean absolute percentage error (MAPE) of 0.25 in regression tasks. These experimental results demonstrate TG-Mamba’s exceptional performance in TMB prediction, highlighting its potential to extend the benefits of immunotherapy to a broader population of lung cancer patients. The code for our model and the experimental data can be obtained at <span><span>https://github.com/ukeLin/TG-Mamba</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"124 ","pages":"Article 102626"},"PeriodicalIF":4.9,"publicationDate":"2025-08-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144858140","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Prediction of hematoma changes in spontaneous intracerebral hemorrhage using a Transformer-based generative adversarial network to generate follow-up CT images","authors":"Changfeng Feng ,&nbsp;Caiwen Jiang ,&nbsp;Chenxi Hu ,&nbsp;Shuaihang Kong ,&nbsp;Ziyi Ye ,&nbsp;Jing Han ,&nbsp;Kai Zhong ,&nbsp;Tingting Yang ,&nbsp;Hongmei Yin ,&nbsp;Qun Lao ,&nbsp;Zhongxiang Ding ,&nbsp;Dinggang Shen ,&nbsp;Qijun Shen","doi":"10.1016/j.compmedimag.2025.102614","DOIUrl":"10.1016/j.compmedimag.2025.102614","url":null,"abstract":"<div><h3>Purpose</h3><div>To visualize and assess hematoma growth trends by generating follow-up CT images within 24 h based on baseline CT images of spontaneous intracerebral hemorrhage (sICH) using Transformer-integrated Generative Adversarial Networks (GAN).</div></div><div><h3>Methods</h3><div>Patients with sICH were retrospectively recruited from two medical centers. The imaging data included baseline non-contrast CT scans taken after onset and follow-up imaging within 24 h. In the test set, the peak signal-to-noise ratio (PSNR) and the structural similarity index measure (SSIM) were utilized to quantitatively assess the quality of the predicted images. Pearson’s correlation analysis was performed to assess the agreement of semantic features and geometric properties of hematomas between true follow-up CT images and the predicted images. The consistency of hematoma expansion prediction between true and generated images was further examined.</div></div><div><h3>Results</h3><div>The PSNR of the predicted images was 26.73 ± 1.11, and the SSIM was 91.23 ± 1.10. The Pearson correlation coefficients (r) with 95 % confidence intervals (CI) for irregularity, satellite sign number, intraventricular or subarachnoid hemorrhage, midline shift, edema expansion, mean CT value, maximum cross-sectional area, and hematoma volume between the predicted and true follow-up images were as follows: 0.94 (0.91, 0.96), 0.87 (0.81, 0.91), 0.86 (0.80, 0.91), 0.89 (0.84, 0.92), 0.91 (0.87, 0.94), 0.78(0.68, 0.84), 0.94(0.91, 0.96), and 0.94 (0.91, 0.96), respectively. The correlation coefficient (r) for predicting hematoma expansion between predicted and true follow-up images was 0.86 (95 % CI: 0.79, 0.90; <em>P</em> &lt; 0.001).</div></div><div><h3>Conclusions</h3><div>The model constructed using a GAN integrated with Transformer modules can accurately visualize early hematoma changes in sICH.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"124 ","pages":"Article 102614"},"PeriodicalIF":4.9,"publicationDate":"2025-08-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144831352","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"FGDN: A Federated Graph Convolutional Network framework for multi-site major depression disorder diagnosis","authors":"Chun Liu ,&nbsp;Shengchang Shan ,&nbsp;Xinshun Ding ,&nbsp;Huan Wang ,&nbsp;Zhuqing Jiao","doi":"10.1016/j.compmedimag.2025.102612","DOIUrl":"10.1016/j.compmedimag.2025.102612","url":null,"abstract":"<div><div>The vast amount of healthcare data is characterized by its diversity, dynamic nature, and large scale. It is a challenge that directly training a Graph Convolutional Neural Network (GCN) in a multi-site dataset poses to protecting the privacy of Major Depressive Disorder (MDD) patients. Federated learning enables the training of a global model without the need to share data. However, some previous methods overlook the potential value of non-image information, such as gender, age, education years, and site information. Multi-site datasets often exhibit the problem of Non-Independent and Identically Distributed (Non-IID) data, which leads to the loss of edge information across local models, ultimately weakening the generalization ability of the federated learning models. Accordingly, we propose a Federated Graph Convolutional Network framework with Dual Graph Attention Network (FGDN) for multi-site MDD diagnosis. Specifically, both linear and nonlinear information are extracted from the functional connectivity matrix via different correlation measures. A Dual Graph Attention Network (DGAT) module is designed to capture complementary information between these two types. Then a Federated Graph Convolutional Network (FedGCN) module is introduced to address the issue of missing edge information across local models. It allows each local model to receive aggregated feature information from neighboring nodes of other local models. Additionally, the privacy of patients is protected with fully homomorphic encryption. The experimental results demonstrate that FGDN achieves a classification accuracy of 61.8% on 841 subjects from three different sites, and outperforms some recent centralized learning frameworks and federated learning frameworks. This proves it fully mines the feature information in brain functional connectivity, alleviates the information loss caused by Non-IID data, and secures the healthcare data.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"124 ","pages":"Article 102612"},"PeriodicalIF":4.9,"publicationDate":"2025-08-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144809902","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"AI-driven multi-modal framework for prognostic modeling in glioblastoma: Enhancing clinical decision support","authors":"Zihan Zhao ,&nbsp;Nguyen Quoc Khanh Le ,&nbsp;Matthew Chin Heng Chua","doi":"10.1016/j.compmedimag.2025.102628","DOIUrl":"10.1016/j.compmedimag.2025.102628","url":null,"abstract":"<div><h3>Objective</h3><div>Glioblastoma (GBM) is the most aggressive malignant brain tumor, associated with poor prognosis and limited therapeutic options. Accurate prognostic modeling is essential for guiding personalized treatment strategies. However, existing models often rely on single-modality data, limiting their ability to capture the complex molecular and histopathological heterogeneity of GBM. This study proposes an AI-driven, multi-modal framework for clinical decision support, encompassing both early triage and prognostic evaluation stages. A Vision Transformer (ViT) is first employed to classify tumor grades (WHO grades 2–4) using radiological images. Subsequently, an attention-based deep learning model integrates histopathological and transcriptomic data to improve risk stratification and inform treatment planning.</div></div><div><h3>Methods</h3><div>The ViT model was trained on FLAIR MRI scans from the UCSF-PDGM dataset to perform tumor grading during the early triage phase. For prognostic modeling, whole-slide histopathological images and RNA sequencing profiles were integrated using an attention-based deep learning architecture. Model performance was evaluated using the area under the curve (AUC), concordance index (C-index), and Kaplan–Meier survival analysis across independent CPTAC-GBM and TCGA-GBM cohorts.</div></div><div><h3>Results</h3><div>The ViT model achieved F1-scores exceeding 0.89 across all WHO tumor grades. The multi-modal model significantly outperformed single-modality baselines, demonstrating higher C-index values and superior prognostic accuracy. Kaplan–Meier analysis revealed statistically significant differences (p &lt; 0.0001) between high- and low-risk patient groups.</div></div><div><h3>Conclusion</h3><div>This AI-enabled, multi-modal framework improves clinical decision support in GBM by enabling accurate risk stratification and treatment planning. The integration of radiological imaging, histopathology, and transcriptomics offers a comprehensive and personalized approach to GBM prognosis.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"124 ","pages":"Article 102628"},"PeriodicalIF":4.9,"publicationDate":"2025-08-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144831355","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Lightweight wavelet convolutional network for guidewire segmentation","authors":"Guifang Zhang ,&nbsp;Dingyue Liu ,&nbsp;Zhe Ji ,&nbsp;Bin Xie ,&nbsp;Qihan Chen","doi":"10.1016/j.compmedimag.2025.102618","DOIUrl":"10.1016/j.compmedimag.2025.102618","url":null,"abstract":"<div><div>Accurate guidewire segmentation is crucial for the success of vascular interventional procedures. Existing methods rely on a large number of parameters, making it difficult to balance performance and model size. In addition, the difficulty of collecting dual guidewire data poses constraints on the training of dual guidewire segmentation models, making dual guidewire segmentation a challenging task. This study aims to propose an efficient and robust lightweight method for accurate segmentation of single and dual guidewire in X-ray fluoroscopy sequences, while overcoming the challenges caused by data scarcity and model complexity. To this end, we propose a lightweight wavelet convolutional network (WT-CMUNeXt) for guidewire segmentation. WT-CMUNeXt integrates wavelet convolution and channel attention mechanisms, enabling efficient extraction of multi-frequency features while minimizing computational complexity. Additionally, a dual guidewire data augmentation algorithm is designed that synthesizes dual guidewire data from single guidewire data to expand the guidewire dataset. Experimental results on multiple patient sequences demonstrate that the proposed WT-CMUNeXt achieves state-of-the-art performance in the single guidewire segmentation task, with an average F1 score of 0.9048 and an average IoU of 0.8284 in most cases. For the more challenging dual guidewire segmentation task, our method also achieved a strong performance with an F1 score of 0.8668, outperforming all other methods except nnUNet. While also maintaining a minimal model size with only 3.26 M parameters and a low computational cost of 2.99 GFLOPs, making it a practical solution for real-time deployment in clinical guidewire segmentation tasks. Our code and datasets are available at: <span><span>https://github.com/pikopico/WT-CMUNeXt</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"124 ","pages":"Article 102618"},"PeriodicalIF":4.9,"publicationDate":"2025-08-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144860896","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"DeepHybrid-CNN: A hybrid approach for pre-processing of skin cancer images","authors":"Sonam Khattar ,&nbsp;Ravinder Kaur ,&nbsp;Abhishek Kumar","doi":"10.1016/j.compmedimag.2025.102611","DOIUrl":"10.1016/j.compmedimag.2025.102611","url":null,"abstract":"<div><div>In the current technological era, digital imaging is ubiquitous, and it serves a crucial purpose in the realm of medical research. Skin cancer is one of the most common types of cancer, and its early diagnosis is essential to reduce the mortality rates. An impediment that must be overcome to identify essential image details for computerized cancer segmentation effectively and classification is the presence of a variety of noise types, as well as hair in the acquired skin lesion images. Pre-processing of skin cancer images is a crucial but difficult challenge in developing highly effective computerized diagnosis systems. The existence of fine lines and low picture quality in the skin images presents a substantial challenge in precisely defining characteristics for automated cancer classification. The purpose of this research is to propose a novel hybrid technique by employing a fusion of Anisotropic Intensity Hair Removal (AI-HR), a Gaussian Filter (GF), and a deep learning based residual convolutional neural network (Deep Residual CNN) to improve the quality of dermoscopic images for performing correct diagnostic tasks. The experimental results revealed that the proposed technique successfully removes hairs and noise from dermoscopic images, resulting in better visibility of details in terms of different evaluation metrics. Further, investigations were carried out to observe how the proposed pre-processing techniques help to enhance the segmentation and classification performance for the diagnosis of skin cancer images. The experimental results revealed that there was an enhancement in the segmentation and classification results by utilizing the proposed hybrid pre-processing technique. Also, the recommended method outperforms the current state-of-the-art pre-processing techniques in the field of skin cancer diagnosis, as shown by the findings of investigations carried out on the HAM10000 dataset. The results revealed that the methodology was superior in both subjective and objective evaluations and has the potential to be deployed in real-time clinical scenarios.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"124 ","pages":"Article 102611"},"PeriodicalIF":4.9,"publicationDate":"2025-08-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144813850","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Surgical augmented reality registration methods: A review from traditional to deep learning approaches","authors":"Alberto Neri ,&nbsp;Veronica Penza ,&nbsp;Chiara Baldini ,&nbsp;Leonardo S. Mattos","doi":"10.1016/j.compmedimag.2025.102616","DOIUrl":"10.1016/j.compmedimag.2025.102616","url":null,"abstract":"<div><div>Augmented Reality (AR) has gained significant interest within the research community in the past two decades. In surgery, AR overlays critical information directly onto the surgeon’s visual field, thus enhancing situational awareness by providing navigation guidance and contributing to safer, more precise, and more efficient surgical interventions. This review examines registration methods suitable for laparoscopic scenarios involving a pre-operative 3D model and an intra-operative 2D or 3D video, both in rigid and non-rigid conditions. We started analysing traditional methods, which do not include Deep Learning (DL). However, in recent years, DL has revolutionized the landscape of computer vision in many tasks. So, we investigated these methods in our scope, identifying two main categories: hybrid DL-enhanced methods and DL point cloud registration methods. The former applies DL across different stages of the traditional surface-based methods to enhance their performances. The latter uses end-to-end approaches to estimate the transformation matrix between two input point clouds. For each category discussed, we highlight both the strengths and weaknesses associated with the challenges of surgical AR to aid comprehension, even for those less familiar with the topic.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"124 ","pages":"Article 102616"},"PeriodicalIF":4.9,"publicationDate":"2025-08-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144866517","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"IgCONDA-PET: Weakly-supervised PET anomaly detection using implicitly-guided attention-conditional counterfactual diffusion modeling — a multi-center, multi-cancer, and multi-tracer study","authors":"Shadab Ahamed ,&nbsp;Arman Rahmim","doi":"10.1016/j.compmedimag.2025.102615","DOIUrl":"10.1016/j.compmedimag.2025.102615","url":null,"abstract":"<div><div>Minimizing the need for pixel-level annotated data to train PET lesion detection and segmentation networks is highly desired and can be transformative, given time and cost constraints associated with expert annotations. Current unsupervised or weakly-supervised anomaly detection methods rely on autoencoder or generative adversarial networks (GANs) trained only on healthy data. While these approaches reduce annotation dependency, GAN-based methods are notably more challenging to train than non-GAN alternatives (such as autoencoders) due to issues such as the simultaneous optimization of two competing networks, mode collapse, and training instability. In this paper, we present the weakly-supervised <strong>I</strong>mplicitly <strong>g</strong>uided <strong>CO</strong>u<strong>N</strong>terfactual diffusion model for <strong>D</strong>etecting <strong>A</strong>nomalies in <strong>PET</strong> images (IgCONDA-PET). The solution is developed and validated using PET scans from six retrospective cohorts consisting of a total of 2652 cases (multi-cancer, multi-tracer) containing both local and public datasets (spanning multiple centers). The training is conditioned on image class labels (healthy vs. unhealthy) via attention modules, and we employ implicit diffusion guidance. We perform counterfactual generation which facilitates “unhealthy-to-healthy” domain translation by generating a synthetic, healthy version of an unhealthy input image, enabling the detection of anomalies through the calculated differences. The performance of our method was compared against several other deep learning based weakly-supervised or unsupervised methods as well as traditional methods like 41% SUV<span><math><msub><mrow></mrow><mrow><mtext>max</mtext></mrow></msub></math></span> thresholding. We also highlight the importance of incorporating attention modules in our network for the detection of small anomalies. The code is publicly available at: <span><span>https://github.com/ahxmeds/IgCONDA-PET.git</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50631,"journal":{"name":"Computerized Medical Imaging and Graphics","volume":"124 ","pages":"Article 102615"},"PeriodicalIF":4.9,"publicationDate":"2025-07-31","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144766922","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}